METHODS FOR MEDIATING FACTORS IN SOCIAL EPIDEMIOLOGY

University Of North Carolina Chapel Hill, Chapel Hill NC

Investigators

Linked publications & trials

Abstract

DESCRIPTION: Disparities in health status by socioeconomic factors are among the oldest and most widely corroborated epidemiologic observations. There has been a recent resurgence of interest in "social epidemiology," and an active research program now seeks to identify the etiologic pathways through which these effects are transmitted. The standard methodologic approach for etiologic inferences concerning a social factor is to decompose effects by contrasting two adjusted effect estimates: one adjusted for potential confounders, and one adjusted for the same potential confounders plus one or more variables hypothesized to lie on the pathway through which the social factor exerts its effect. This contrast is typically used to distinguish indirect effects, through the specified intervening variables, from direct effects, transmitted via unspecified pathways. If control of hypothetical causal intermediates greatly attenuates an estimated social factor effect, it is generally inferred that the effect is mediated primarily through pathways involving these quantities; minimal attenuation is interpreted as evidence that other pathways predominate. These mechanistic inferences then inform policy decisions concerning the utility of potential interventions. Although the decomposition approach is now widespread in social epidemiology, its general validity is uncertain. Previous methodologic work suggests that the strategy may have substantial limitations which have never been studied in the social epidemiologic context. This proposed project will investigate the common strategy in social epidemiology of making mechanistic inferences by controlling analytically for hypothetical causal intermediates "as though they were confounders." The assumptions and conditions necessary to support valid etiologic inferences using this approach will be evaluated formally and through simulation exercises, and the sensitivity of conventional statistical models to plausible violations of the necessary assumptions will be quantified using social factor and outcome data from the National Longitudinal Mortality Study and the Panel Study of Income Dynamics. Analytic traditions in epidemiology and the social sciences will be examined in contrast with modern alternatives, including Pearl's nonparametric structural equations and Robins' g-computation algorithm. This will help to clarify conditions under which standard effect decomposition may be useful for social factors, and when alternative approaches may De preferred. This will thereby provide a basis for more substantial advancement in social epidemiology, and thus for efficacious policies toward the successful elimination of health disparities in accordance with current national policy directives.